Skip to content

v0.4.0 — SmolVLA tasks and an MCP Blender director

Choose a tag to compare

View all tags
@noteflowai noteflowai released this 12 Sep 03:39
v0.4.0
38e0bda
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Robot Reel 0.4.0 makes Physical AI runs easier to watch, inspect and reuse.

  • SmolVLA task replay: an actual CPU policy rollout in LIBERO, with synchronized scene/wrist cameras, every applied action, measured robot states and the task outcome. The published seeded run completed the bowl-to-plate task after 76 actions.
  • MCP agent director: inspect recorded events, create a validated storyboard, and build an editable Blender film with four cameras, captions and sample-preserving slow motion. The included film retains 180 source samples across 210 frames.
  • Newton → OpenUSD → Blender: the existing CPU rigid-body replay and native transform checks are included in this release.

Try the demos:

Downloads below include the complete offline VLA episode and the editable director project. Preserve the media attribution included in the episode. Reproduction instructions and the Chinese README are in the repository.

Validation: 46 Python tests, 18 browser checks, actual MCP stdio calls, real CPU policy inference, and native Blender checks of all 420 directed vehicle samples. Videos and downloads are tied to their recorded traces by checksums. The VLA result is one simulation rollout, not a benchmark or hardware validation; the website plays recorded examples.

Assets 7
Loading